S.V. Astashkin, F.A. Sukochev and C.P. Wong- Disjointification of Martingale Differences and...
Transcript of S.V. Astashkin, F.A. Sukochev and C.P. Wong- Disjointification of Martingale Differences and...
-
8/3/2019 S.V. Astashkin, F.A. Sukochev and C.P. Wong- Disjointification of Martingale Differences and Conditionally Independ
1/32
DISJOINTIFICATION OF MARTINGALE DIFFERENCES AND
CONDITIONALLY INDEPENDENT RANDOM VARIABLES
WITH SOME APPLICATIONS.
S. V. ASTASHKIN, F. A. SUKOCHEV, AND C. P. WONG
Abstract. In this paper, disjointification inequalities are proven for arbi-
trary martingale difference sequences and conditionally independent random
variables of the form {fk(s)xk(t)}nk=1
, where fks are independent and xks
are arbitrary random variables from a symmetric space X on [0, 1]. The main
results show that the form of these inequalities depends on which side from
L2 the space X lies. The disjointification inequalities obtained allow us to
compare norms of sums of martingale differences and nonnegative random
variables with the norms of sums of their independent copies. The latter re-
sults can be treated as an extension of modular inequalities proved earlier by
de la Pena and Hitczenko to the setting of symmetric spaces. Moreover, using
these results simplifies the proofs of some modular inequalities.
1. Introduction
In 1970, H. P. Rosenthal proved a remarkable inequality [29] from which it follows
that for sequences of independent mean zero random variables in Lp[0, 1], p 2,the mapping fk fk, where fk(t) := fk(t k + 1)[k1,k)(t) (t > 0), extendsto an isomorphism between the closed linear span [fk]
k=1 (in Lp[0, 1]) and the
closed linear span [fk]k=1 (in Lp[0,) L2[0,)). A significant generalization
of this disjointification inequality to the class of symmetric spaces X on [0, 1] is
due to W. B. Johnson and G. Schechtman [21]. In particular, they introduced the
symmetric space Zp
X
on [0,
) (Note: Our notation differs from that used in [21])
which can be defined as the space of all functions f L1[0,) + L[0,) suchthat fZpX :=
f[0,1]X + f[1,)p < , where f is the non-increasingrearrangement of f (see details in Section 2). In their paper [21], they showed
that any sequence {fk}k=1 of independent mean zero random variables in X isResearch was partially supported by the ARC and the RFBR grant no. 10-01-00077.
Research was partially supported by the ARC.
1
-
8/3/2019 S.V. Astashkin, F.A. Sukochev and C.P. Wong- Disjointification of Martingale Differences and Conditionally Independ
2/32
2 S. V. ASTASHKIN, F. A. SUKOCHEV, AND C. P. WONG
equivalent to the sequence of its disjoint translates {fk}k=1 in Z2X provided that Xcontains an Lp-space for some p t}) = ({s [0, ) : |y(s)| > t}) (t > 0)
(we will say in this case that the functions |x| and |y| are equimeasurable)and y X, then x X and xX = yX .
Also, if X is a symmetric space on I = [0, 1] then X(I I) is the correspondingsymmetric space on the square with the norm xX(II) = xX . Here, x denotes
-
8/3/2019 S.V. Astashkin, F.A. Sukochev and C.P. Wong- Disjointification of Martingale Differences and Conditionally Independ
4/32
4 S. V. ASTASHKIN, F. A. SUKOCHEV, AND C. P. WONG
the non-increasing, left-continuous rearrangement of x, which is equimeasurable
with|x|
and given by
x(t) = inf{ > 0 : 2({s : |x(s)| > }) < t}, t > 0
where 2 is the 2-dimensional Lebesgue measure on I I.Important examples of symmetric spaces are Lorentz (including Lp,q spaces) and
Orlicz spaces, which will be denoted by and LM respectively. For more detailed
information about these spaces and other basic properties of symmetric spaces,
refer to [25, 23, 12].
The following symmetric space ZpX introduced in [19] (see also [21]) will play an
important role in this paper.
Definition 2.2. For an arbitrary symmetric space X on [0, 1] and any p [1,],we define the function space ZpX on [0,) by
ZpX := {f L1[0,) + L[0,) : fZpX
:=f[0,1]X + f[1,)p < }.
The Kothe dual X of a symmetric space X on the interval [0, ) consists of all
measurable functions y for which
yX = sup0|x(t)y(t)| dt : x X, xX 1 < .
Basic properties of Kothe duality can be found in [25] and [23].
We say that X has
(1) order continuous normif from {xk}k=1 X, xk k 0 a.e. on [0, ) it followsthat xkX 0;
(2) order semi-continuous norm if from {xk}k=1 X, x X and xk x a.e.on [0, ), it follows that xX lim infn xnX ;
(3) the Fatou property, if from
{xk
}k=1
X, xk
x a.e. on [0, ) and
supk xkX < , it follows that x X and xX lim infk xkX .For a concise summary of how these properties relate to the properties ofX see [9,
Section 2].
Throughout the paper, we will denote the set of all 1-interpolation spaces (see
for e.g. [11, 23, 25]) between A and B by I(A, B). For a sequence of functions
{fk}k=1 X(I), denote their disjoint translates fk(t) := fk(tk+1)[k1,k)(t), t >
-
8/3/2019 S.V. Astashkin, F.A. Sukochev and C.P. Wong- Disjointification of Martingale Differences and Conditionally Independ
5/32
5
0. Analogously, if{fk(t)xk(s)}k=1 X(I I), then fk(t)xk(s) := (fkxk)(t k +1)[k1,k)(t), t > 0.
Finally, if X is a symmetric function space on [0, 1] or [0,), the Boyd indicespX and qX are defined by
pX = lims
log s
log s = sups>1log s
log sand
qX = lims0+
log s
log s = sup0
-
8/3/2019 S.V. Astashkin, F.A. Sukochev and C.P. Wong- Disjointification of Martingale Differences and Conditionally Independ
6/32
6 S. V. ASTASHKIN, F. A. SUKOCHEV, AND C. P. WONG
There is an operator K defined on S([0, 1], ) which is closely linked to theKruglov property (see [6, 7, 8]). Throughout this paper (G, ) :=
k=0
([0, 1], k),
where k is the Lebesgue measure on [0, 1] for all k 0.
Definition 2.4. Let {Am} be a sequence of pairwise disjoint subsets of [0,1] with(Am) = 1/(e m!), m N. Given an f S([0, 1], ), we set
Kf(0, 1, . . . ) :=
m=1
mj=1
f(j)Am (0)
Since we work only with symmetric spaces, the main object of interest is the
distribution of the function Kf. Hence we can also consider the following equivalentform of the operator K.
If f S([0, 1], ) and {fm,j}mj=1, m N is a sequence of measurable functions on[0, 1] such that for every m N, fm,1, fm,2, . . . , f m,m, Am are independent randomvariables with Ffm,j = Ff, j = 1, 2, . . . , m, then we write
(2.1) Kf(x) :=
m=1
mj=1
fm,j(x)Am (x), x [0, 1].
Since for all f S([0, 1], ) and t R,
{(0, 1, . . . )
G :
Kf(0, 1, . . . ) > t
}=
{s
[0, 1] :
Kf(s) > t
},
we can regard K as an operator acting from S([0, 1], ) to S([0, 1], ).It was shown in [6] that if X is a symmetric space on [0, 1], then the operator
K maps X boundedly into itself if and only if X K. In fact, in [6, 7, 8, 9], theoperator K is shown to play an important role in estimating the norms of sums ofindependent random variables by the norms of the sums of their disjoint translates.
In Section 4, we show how an extended version of the operator K plays a similarrole for conditionally independent random variables.
To do so, we will need to move into Banach function spaces with mixed norm
(see, for example, [22, 11.1, p. 400]). Let X be a Banach lattice and Y be a Banachlattice with order semi-continuous norm on I = [0, 1]. The space with mixed norm
X[Y] consists of all measurable on the square I I functions x(s, t) satisfying theconditions: 1) for a.e. s I the function x(s, ) Y; 2) the function x(s) =x(s, )Y X. Then X[Y] endowed with the norm xX[Y] = xX is a Banachlattice on I I.
-
8/3/2019 S.V. Astashkin, F.A. Sukochev and C.P. Wong- Disjointification of Martingale Differences and Conditionally Independ
7/32
7
3. Disjointification of martingale differences
In this section we establish one-sided disjointification inequalities for martingaledifference sequences (mds) in symmetric spaces. The main results, Theorem 3.1
and 3.5, give necessary and sufficient conditions for the right-hand-side and left-
hand-side disjointification inequalities respectively. We will also show that the form
of these inequalities essentially depends on which side from L2 the space X lies.
Let X be a symmetric space on [0, 1]. Denote by X(l2) the set of all sequences
{xk}k=1 of functions from X such that
{xk
}X(l2)
:= sup
n=1,2,...
n
k=1
x2k1/2
X < .
The closed subspace of the space X(l2) generated by the set of all eventually van-
ishing sequences {xk} we will denote by X(l2).Before stating the main result for the right-hand-side disjointification inequality
for mds (see (3.2)), we need to define a linear operator which acts on the space
S(0,) into the space of sequences of functions from S(0, 1) by
(3.1) Bx(t) := {x(t + k 1)}k=1 (0 t 1).
It turns out that the boundedness of the operator B gives a necessary and sufficient
condition for the right-hand disjointification inequality .
Theorem 3.1. Let X be a symmetric space on [0, 1] such that qX < . Thefollowing conditions are equivalent:
(i) B is bounded from Z2X intoX(l2);
(ii) there exists C > 0 such that for any mds {dk}nk=1 X we have that
(3.2) maxk=1,2,...,n k
i=1
diX C
n
i=1
diZ2X
;
(iii) there exists C > 0 such that for any sequence {xk}nk=1 X we have that
(3.3) n
i=1
ri(t)xi(s)X(II)
C n
i=1
xi
Z2X
,
where {ri} are the Rademacher functions, i.e., ri(t) = sign(sin2it) (i =1, 2, . . . ) for t [0, 1].
-
8/3/2019 S.V. Astashkin, F.A. Sukochev and C.P. Wong- Disjointification of Martingale Differences and Conditionally Independ
8/32
8 S. V. ASTASHKIN, F. A. SUKOCHEV, AND C. P. WONG
Proof. Since (i) holds, we have that
(3.4) n
i=1
d2i1/2
X=B( n
i=1
di)X(l2)
BZ2XX(l2) n
i=1
diZ2X
for any mds {dk}nk=1 X. The assumption qX < then allows us to apply [20,Theorem 3] to get
maxk=1,2,...,n
ki=1
di
X C1
ni=1
d2i
1/2X
for any mds {dk}nk=1 X. Combining this with (3.4) we obtain (3.2).Since the Rademacher functions {ri} are independent, then the implication
(ii)
(iii) is trivial. So it remains to prove only implication (iii)
(i). By a well-
known consequence of the Khintchine inequality [25, 2.d.1], for every symmetric
space X there is a constant c > 0 such that
ni=1
ri(t)xi(s)X(II)
c n
i=1
xi21/2
X.
Therefore, from (3.3), it follows that
ni=1
xi21/2
X C
ni=1
xi
Z2X
for an arbitrary sequence
{xk}nk=1
X. It is clear that this is equivalent to (i).
The following result is an easy consequence of Theorem 3.1 and [4, Lemma 3.5]
which asserts that if X is a symmetric space on [0, 1], X I(L1, L2), then theoperator B is bounded from Z2X into
X(l2).
Corollary 3.2. Let X be a symmetric space on [0, 1], X I(L1, L2). Then thereexists a constant C > 0 such that (3.2) holds for every mds {dk}nk=1 X.
In the following corollary, we consider a special case of mds, which is related to
the conditionally independent random variables we study in Section 5.
Corollary 3.3. Let X be a symmetric space on [0, 1], X I(L1, L2). Then thereexists a constant C > 0 such that for every sequence {fk}nk=1 X of mean zeroindependent functions and arbitrary sequence {xk}nk=1 X we have that
(3.5) max
k=1,2,...,n
ki=1
fi(s)xi(t)
X(II) C
ni=1
fi(s)xi(t)Z2X
.
-
8/3/2019 S.V. Astashkin, F.A. Sukochev and C.P. Wong- Disjointification of Martingale Differences and Conditionally Independ
9/32
9
Proof. It is sufficient to note that the sequence {fi(s)xi(t)}ni=1 is an mds withrespect to the increasing sequence of algebras,
{k}n
k=1
, of subsets of the square
I I where k is generated by the functions {fi(s)xi(t)}ki=1 (k = 1, . . . , n) andapply the previous corollary.
Let us show that any of the conditions of Theorem 3.1 implies that X L2.
Theorem 3.4. Let X be a symmetric space on [0, 1] that is separable or has the
Fatou property. If the operator B is bounded from Z2X intoX(l2), then X L2.
Proof. Without loss of generality, we may (and shall) assume that [0,1]X = 1.
Suppose, on the contrary, that X L2. Since X either has the Fatou propertyor is separable, it follows that for an arbitrary M > 0 there is a step function
x(t) =m
j=1 ajEj (t) such that the sets Ej [0, 1] are pairwise disjoint, aj 0(j = 1, . . . , m), xL2 = 1, and xX M + 1. Denote y(t) =
mj=1[aj ]Ej (t),
where [a] is the integer part of a real number a. Then x = y + z, where |z(t)| 1.Therefore, by the previous inequality,
(3.6) yX M.
Setting yi(t) := min(y(t)2, i) and fi = yi yi1 (i = 1, 2, . . . ), where y0 = 0, weobtain that y(t)2 =
li=1 fi(t) for some l 1. Since fis are indicator functions of
some measurable subsets of [0, 1], we see that f :=l
i=1 fi is the indicator function
of a measurable subset of (0,) whose Lebesgue measure equals mj=1[aj ]2(Ej).Hence,
fZ2X = f[0,1]X + f[1,)L2 1 + m
j=1
[aj ]2(Ej)
1/2 1 + xL2 2.
On the other hand, sincel
i=1 fi(t) =l
i=1 fi(t)2, t > 0, we have
y(t) = m
j=1[aj ]
2
Ej (t)1/2
= l
i=1fi(t)
21/2.
Thus, by (3.6),
BfX(l2) = l
i=1
fi(t)21/2
X= yX M.
Since M is arbitrarily large and fZ2X 2 independently of M, this contradictsour assumption of the boundedness of B and the proof is complete.
-
8/3/2019 S.V. Astashkin, F.A. Sukochev and C.P. Wong- Disjointification of Martingale Differences and Conditionally Independ
10/32
10 S. V. ASTASHKIN, F. A. SUKOCHEV, AND C. P. WONG
Now consider the converse of inequalities (3.2) and (3.5). In this case, the bound-
edness of the operator A defined below provides a necessary and sufficient condition
for the inequalities to hold.
Let X be a symmetric space on [0, 1]. Define on X(l2) the following linear oper-
ator acting into the set S(0,):
A({xk})(u) :=k=1
xk(u k + 1)[k1,k)(u).
The proof of Theorem 3.5 is very similar to the proof of Theorem 3.1 and is
therefore, omitted.
Theorem 3.5. Let X be a symmetric space on [0, 1] such that qX < . Thefollowing conditions are equivalent:
(i) the operator A is bounded from X(l2) into Z2X ;
(ii) there exists C > 0 such that for any mds {dk}nk=1 X we have that
(3.7) n
i=1
di
Z2X
C max
k=1,2,...,n
ki=1
di
X
;
(iii) there exists C > 0 such that for any sequence {xk}nk=1 X we have that
(3.8) n
i=1
xi
Z2X
C n
i=1
ri(t)xi(s)X(II)
.
The following corollary is the analogue of Corollaries 3.2 and 3.3 for the left-
hand-side inequality.
Corollary 3.6. Let X be a symmetric space on [0, 1] with order semi-continuous
norm such that qX < and X I(L2, L). Then there exists a constant C > 0such that (3.7) holds for every mds {dk}nk=1 X.
In particular, for every sequence {fk}nk=1 X of mean zero independent func-tions and arbitrary sequence {xk}nk=1 X we have that
(3.9) n
i=1
fi(t)xi(s)Z2X
C max
k=1,2,...,n
ki=1
fi(t)xi(s)
X(II).
Proof. By Theorem 3.5, it suffices to show that the operator A is bounded from
X(l2) into Z2X for spaces X satisfying the conditions above.
-
8/3/2019 S.V. Astashkin, F.A. Sukochev and C.P. Wong- Disjointification of Martingale Differences and Conditionally Independ
11/32
11
Firstly, it is easy to check that the space Z2X has order semi-continuous norm as
X does. Therefore, from the equality (Z2X) = Z2X (see for e.g. [4, Lemma 3.3]), it
follows that
(3.10) A({xk})Z2X = supy
Z2X1
0
A({xk})(u)y(u) du
for every {xk}k=1 X(l2).Moreover, for any {xk}k=1 X(l2) and any y Z2X we have
(3.11)
0
A({xk})(u)y(u) du =k=1
10
(By)k(u)xk(u) du
which follows from0
A({xk})(u)y(u) du =
0
k=1
xk(u k + 1)[k1,k)(u)y(u) du
=
k=1
kk1
xk(u k + 1)y(u) du
=
k=1
10
y(u + k 1)xk(u) du
=k=1
10
(By)k(u)xk(u) du.
Since (L2, L) is a K-monotone couple [13, 4.4.38] and L2 + L
= L2 + L1
is separable, it follows from [27, Theorem 4.1] that X I(L1, L2). Hence, by [4,Lemma 3.5], the operator B is bounded from Z2X into
X(l2). Then, using (3.10)
and (3.11), we obtain that
A{xk}Z2X
supy
Z2X1
k=1
10
(By)k(u)xk(u) du
supy
Z2X1
supn=1,2,...
nk=1
(By)2k
1/2X
nk=1
x2k
1/2X
supyZ2X1 ByX(l2){xk}X(l2) = B{xk}X(l2).
Therefore, A is bounded from X(l2) into Z2X and the proof is complete.
Next, we show that any of the conditions of Theorem 3.5 implies that X L2.
Theorem 3.7. Let X be a symmetric space on [0, 1] that is separable or has the
Fatou property. If the operator A is bounded from X(l2) into Z2X , then X L2.
-
8/3/2019 S.V. Astashkin, F.A. Sukochev and C.P. Wong- Disjointification of Martingale Differences and Conditionally Independ
12/32
12 S. V. ASTASHKIN, F. A. SUKOCHEV, AND C. P. WONG
Proof. We claim that the operator B is bounded from Z2X intoX(l2). In fact, for
any y
Z2X and n
N
nk=1
(By)2k
1/2X
= supxX1
10
nk=1
(By)k(u)21/2
x(u) du.
It is clear that there are measurable functions zk(t) (k = 1, 2, . . . , n) such thatnk=1 z
2k(u) = 1 (0 u 1) and n
k=1
(By)k(u)21/2
=n
k=1
(By)k(u)zk(u) (0 u 1).
If x X and xk = xzk (k = 1, 2, . . . , n) and xk = 0 (k > n), then {xk} X(l2)and
{xk}
X(l2) =
xX . Therefore, n
k=1
(By)2k
1/2X
sup{xk}X(l2)1
nk=1
10
(By)k(u)xk(u) du.
Hence, by hypothesis and equality (3.11), we have that
ByX(l2) = supn=1,2,... n
k=1
(By)2k
1/2X
A yZ2X
,
and our claim is proved.
Therefore, we can apply Theorem 3.4 and conclude that X L2, or equivalently,X
L2. This implies X
L2 and the proof is complete.
The following corollary is an immediate consequence of Theorems 3.1, 3.4, 3.5
and 3.7. It shows, in contrast to the case of sequences of independent random
variables ([7, Theorem 3.1] and [21, Theorem 1]), that two-sided disjointification
inequalities hold only in L2.
Corollary 3.8. For a given symmetric space X on [0, 1] the following conditions
are equivalent:
(1) for every mds {dk}nk=1 X maxk=1,2,...,n
ki=1
di
X n
i=1
di
Z2X
;
(2) for every sequence {fk}nk=1 X of mean zero independent functions andarbitrary sequence {xk}nk=1 X max
k=1,2,...,n
ki=1
fi(t)xi(s)
X(II) n
i=1
fi(t)xi(s)Z2X
;
-
8/3/2019 S.V. Astashkin, F.A. Sukochev and C.P. Wong- Disjointification of Martingale Differences and Conditionally Independ
13/32
13
(3) for any sequence {xk}nk=1 X
ni=1
ri(t)xi(s)X(II)
ni=1
xiZ2X
;
(4) X = L2 (with equivalence of norms).
Remark 3.9. LetX be the Lorentz space L2,q. It is shown in [16, pp. 154-155] that
in the case when q = 2 neither is the operator B bounded from Z2X into X(l2) noris the operator A bounded from X(l2) into Z
2X . Therefore, for L2,q (q = 2) neither
(3.2) nor (3.7) holds.
Later, we will see that better disjointification results can be proved for condition-
ally independent random variables of a special form. To this end, we first introduce
a modification of the operator K.
4. On the operator K 1X
Definition 4.1. Let E, F and X be Banach function lattices on [0, 1]. Given
a linear operator T : E F and the identity operator 1X on X we define theoperator T 1X : EX FX as acting on all finite combinations of the form
nk=1
fk xk :=n
k=1
fk(s) xk(t), fk E and xk X (k = 1, 2, . . . , n)
as follows:
(T 1X)
nk=1
fk xk
=n
k=1
T(fk) xk.
Our aim is to find the conditions under which the operator T 1X can beextended to the space with mixed norm E[X].
The following result is well known (see [1, Theorem 12.3] for e.g.).
Proposition 4.2. Let E and F be Banach function lattices. If T is a positive
linear operator such that T(E) F, then T is bounded from E into F.
Theorem 4.3. Let E and X be symmetric spaces on [0, 1] that either have the
Fatou property or are separable and E = L, X = L. If T is a positive linearoperator such that T(E) E and 1X is the identity operator on X, then T 1X
-
8/3/2019 S.V. Astashkin, F.A. Sukochev and C.P. Wong- Disjointification of Martingale Differences and Conditionally Independ
14/32
14 S. V. ASTASHKIN, F. A. SUKOCHEV, AND C. P. WONG
can be extended to a bounded operator defined on the whole space E[X] (which will
also be denoted by T
1X) and
T 1XE[X]E[X] TEE .
Proof. Firstly, let u E X i.e. u(s, t) = ni=1 fi(s)xi(t), where n N, fi E, xi X. If y X, then by the linearity of the integral, it follows that
(4.1) (T 1X)1
0
u(s, t)y(t) dt =
10
(T 1X)u(s, t)y(t) dt.
Moreover, for every y X, yX = 1, we have
1
0
u(s, t)y(t) dt u(s, )X .Hence,
T(u(s, )X) =(T 1X)(u(s, )X)
(T 1X)1
0
u(s, t)y(t) dt
(T 1X)
10
u(s, t)y(t) dt
(since T 1X is positive)=
10
(T 1X)u(s, t)y(t) dt
(from (4.1)).
Therefore,
T(u(s, )X) supyX
y=1
1
0
(T 1X)u(s, t)y(t) dt
= (T 1X)u(s, )X (since X has Fatou property or is separable).
Now, applying the monotonicity of the norm of E and Proposition 4.2, we get
(T 1X)uXE T(uX)E TEE uX E = TEE uE[X].
Thus, the operator T 1X acts boundedly from the space E X endowed with
the norm E[X] into E[X] and(4.2) T 1XEXE[X] TEE .
Now we want to extend the operator T1X to the whole space E[X]. IfE andX are separable symmetric spaces, then the set of functions of the form
(4.3)n
i=1
Ai(t)fi(s),
-
8/3/2019 S.V. Astashkin, F.A. Sukochev and C.P. Wong- Disjointification of Martingale Differences and Conditionally Independ
15/32
15
where n N, Ai [0, 1] are pairwise disjoint, and fi L, is dense in the spaceE[X] [22,
11.1, Lemma 2]. Therefore, by (4.2), T
1X can be extended to the
space E[X] so that
T 1XE[X]E[X] TEE .
Let symmetric spaces E and X have the Fatou property. Then, as before, we
extend T 1X to the operator bounded from E0[X0] into E[X], where E0 and X0are the subspaces of the spaces E and X respectively, with order continuous norm
(see also [11, Theorem 3.8]). Hereby, T 1X is defined on L(I I) E0[X0].Let u E[X], u 0. Setting um(s, t) := min(u(s, t), m) (m N), we have that
um u a.e. on I I. Since T is positive, the sequence (T 1X)um increases a.e.Moreover, as was proved, for all m N,
(T 1X)umE[X] TEE umE[X] TEE uE[X] .
Using the definition of a space with mixed norm it is not hard to check that together
with spaces E and X, the space E[X] has the Fatou property. Therefore,
(T 1X)u := limm
(T 1X)um E[X]
and
(4.4) (T 1X)uE[X] TEE uE[X] .
For arbitrary u E[X], we let u = u+ u, where u+ := max(u, 0) and u :=max(u, 0), and set (T 1X)u := (T 1X)u+ (T 1X)u. Then, by (4.4),
(T 1X)uE[X] (T 1X)(|u|)E[X] TEE uE[X] ,
and the proof is complete.
The following result is an immediate consequence of Theorem 4.3 and the fact
that the Kruglov operator, K, is a positive linear operator.
Theorem 4.4. Let E and X be symmetric spaces on [0, 1] that either have the
Fatou property or are separable. IfK(E) E and 1X is the identity operator onX, thenK 1X is bounded in E[X] and
K 1XE[X]E[X] KEE .
-
8/3/2019 S.V. Astashkin, F.A. Sukochev and C.P. Wong- Disjointification of Martingale Differences and Conditionally Independ
16/32
16 S. V. ASTASHKIN, F. A. SUKOCHEV, AND C. P. WONG
Corollary 4.5. If a symmetric space X on I = [0, 1] is such that qX < andeither has the Fatou property or is separable, then
K1X[0,1] : X(I
I)
X(I
I).
Proof. Since the Kruglov operator is bounded in Lp := Lp(I) if 1 p < andLp[Lp] is isometrically isomorphic to the space Lp(I I), from Theorem 4.4, itfollows that K 1Lp[0,1] : Lp(I I) Lp(I I) (1 p < ) and
K 1Lp[0,1]Lp(II)Lp(II) KLpLp .Let now qX < r < . Since X either has the Fatou property or is separable, it
is an interpolation space between L1 and L [23, Theorems 2.4.9 and 2.4.10]. More
specifically, from a onesided version of the Boyd interpolation theorem [3, Theo-
rem 1], it follows that X is an interpolation space between L1 and Lr. Therefore,
K 1X[0,1] is bounded in X(I I), and the proof is complete.
5. Disjointification of conditionally independent random variables
By applying the results of the previous chapter, we are able to prove better results
for conditionally independent random variables of the form {fk(s)xk(t)}nk=1, wherefks are independent and xks are arbitrary, in the case when
(5.1)n
k=1
(supp fk) 1.
The idea of the following theorem appeared in [6] and [7] but was not presented
explicitly. We restate it here as it plays a key role in the proof of the main theorem,
Theorem 5.3. Recalling that we use f to denote the characteristic function of a
random variable f, we have
Theorem 5.1. If a sequence {fk}nk=1 S([0, 1], ) consists of pairwise disjointly
supported functions satisfying condition (5.1), then{Kfk}nk=1 is a sequence of in-
dependent random variables.
Proof. We find the characteristic function of the random vector (Kf1,Kf2, . . . ,Kfn).Note first that
(5.2)
nk=1
(exp(itkfk) 1) = exp(in
k=1
tkfk) 1 (tk R)
-
8/3/2019 S.V. Astashkin, F.A. Sukochev and C.P. Wong- Disjointification of Martingale Differences and Conditionally Independ
17/32
17
which follows from
nk=1
(exp(itkfk) 1) =n
k=1
(exp(itkfk)) n
= exp(in
k=1
tkfk) + n 1 n (since fks are disjointly supported)
= exp(in
k=1
tkfk) 1.
Therefore,
f1,...,fn(t1, . . . , tn) 1 = 1
0 exp(in
k=1
tkfk(x)) 1 dx=
10
nk=1
(exp(itkfk(x)) 1) dx (by (5.2))
=
nk=1
(fk(tk) 1).(5.3)
Finally, this leads to the equality:
Kf1,Kf2,...,Kfn(t1, t2, . . . , tn) = P
nk=1 tkKfk
(1) = K(P
nk=1 tkfk)
(1)
= exp(Pnk=1 tkfk
(1)
1) = exp(f1,...,fn(t1, . . . , tn)
1)
= exp(n
k=1
(fk(tk) 1)) (by (5.3))
=n
k=1
Kfk(tk).
Applying the fact that random variables 1, 2, . . . , n are independent iff
1,2,...,n(t1, t2, . . . , tn) = 1(t1)2(t2) . . . n(tn) for all real t1, t2, . . . tn (see [30,
p. 284] for e.g.), we obtain that the random variables {Kfk}nk=1 are independent.
Next, we will need an auxiliary statement. In what follows, we consider the cubenk=1[0, 1] with usual ndimensional Lebesgue measure n.
Lemma 5.2. Let{fk}nk=1 S([0, 1], ) be a sequence of nonnegative independent functions and{xk}nk=1 S([0, 1], ) be a sequence of arbitrary nonnegative mea-surable functions. Suppose that {hk}nk=1 is a sequence of independent functions on[0, 1] such that hk andKfk are equimeasurable for every k = 1, 2, . . . , n. Then, for
-
8/3/2019 S.V. Astashkin, F.A. Sukochev and C.P. Wong- Disjointification of Martingale Differences and Conditionally Independ
18/32
18 S. V. ASTASHKIN, F. A. SUKOCHEV, AND C. P. WONG
every > 0,
(5.4)
2
(s, t) I I :
nk=1
fk(s)xk(t) >
22
(s, t) I I :
nk=1
hk(s)xk(t) >
2
.
Proof. Since fi 0, by the definition of the Kruglov operator, it follows thatKfi gi, where gi is equimeasurable with the function 1/2fi (i = 1, 2, . . . , n). Inparticular, we have that
(5.5)n
i=1
Kfi(ti)xi(t0) n
i=1
gi(ti)xi(t0).
It is clear that, for every i = 1, 2, . . . , n , the function fi is equimeasurable withthe sum gi + g
i of two disjointly supported functions g
i and g
i , each of which is
equimeasurable with the function gi. Therefore, by (5.5), we have
n+1
(tk)
nk=0 :
ni=1
(gi(ti) + gi (ti))xi(t0) >
2n+1
(tk)
nk=0 :
ni=1
gi(ti)xi(t0) > /2
2n+1
(tk)nk=0 :
ni=1
Kfi(ti)xi(t0) > /2
.
On the other hand, from Fubinis theorem and the assumptions on fi, gi, g
i as well
as hi and Kfi (i N), it follows that, for all > 0,
2
(s, t) I I :
ni=1
fi(s)xi(t) >
= n+1
(tk)
nk=0
nk=0
[0, 1] :ni=1
fi(ti)xi(t0) >
= n+1
(tk)
nk=0 :
ni=1
(gi(ti) + gi (ti))xi(t0) >
and
2(s, t) I I :n
i=1
hi(s)xi(t) > = n+1(tk)nk=0 :
n
i=1
hi(ti)xi(t0) > = n+1
(tk)
nk=0 :
ni=1
Kfi(ti)xi(t0) >
.
Combining these with the previous inequality, we obtain the required result.
The main result of this chapter is following.
-
8/3/2019 S.V. Astashkin, F.A. Sukochev and C.P. Wong- Disjointification of Martingale Differences and Conditionally Independ
19/32
19
Theorem 5.3. Let X be a symmetric space on [0, 1] that either has the Fatou
property or is separable. Suppose qX 0 such
that for every sequence of independent functions {fk}nk=1 X satisfying condition(5.1) and for every sequence of arbitrary measurable functions {xk}nk=1 X wehave that
(5.6)
n
k=1
fk(s)xk(t)
X(II)
C
nk=1
fk(s)xk(t)
X
.
Proof. Without loss of generality, we may (and will) assume that fk 0 and xk 0.Firstly, from Theorem 5.1, it follows that the functions Kfk are independent.
Moreover, it is clear thatK
fk andK
fk are equimeasurable for each k = 1, 2, . . . , n .
Therefore, by [23, Corollary 2.4.2], Lemma 5.2 and Corollary 4.5, we obtain thatn
k=1
fk(s)xk(t)
X(II)
4
nk=1
Kfk(s)xk(t)X(II)
4C
nk=1
fk(s)xk(t)
X(II)
= 4C
nk=1
fk(s)xk(t)
X
,
where C
is the norm of the operator K 1 in the space X(I I).
Remark 5.4. The converse of inequality (5.6) holds in every symmetric space X
that either has the Fatou property or is separable (see the proof of Theorem 6.5 in
the next Section).
The following result shows the necessity of the condition qX < in the lasttheorem.
Theorem 5.5. Suppose that a symmetric space X on [0, 1] has the following prop-
erty: there exists a constant C > 0 such that for every sequence of independent
functions {fk}nk=1 X satisfying condition (5.1) and for every sequence of arbi-trary measurable functions {xk}nk=1 X, inequality (5.6) holds.
Then qX < .
Proof. If qX = , then from Krivines theorem for symmetric spaces (see for e.g.[25, Theorem 2.b.6] or [10, Theorem 4]), it follows that for every integer m, X
-
8/3/2019 S.V. Astashkin, F.A. Sukochev and C.P. Wong- Disjointification of Martingale Differences and Conditionally Independ
20/32
20 S. V. ASTASHKIN, F. A. SUKOCHEV, AND C. P. WONG
contains m pairwise disjointly supported functions {gi}mi=1 having all the samedistribution so that
(5.7)1
2max
i=1,...,m|ai|
mi=1
aigi
2 maxi=1,...,m |ai|for every choice of scalars {ai}mi=1. Obviously, we may assume that gi 0.
Now fix n and consider the aforementioned sequence {gi}mi=1 with m = n2n 2n.Consider the sequence of functions {xk}2nk=1, where xk :=
kn2ni=1 gi. Note that 0
xk xk+1 for k = 1, . . . , 2n1. Moreover, let {fk}2nk=1 be a sequence of independentcopies of the function [0,2n]. We show that, for the sequences {fk(s)xk(t)}2
n
k=1,
the norms on the right-hand side of inequality (5.6) are bounded but the ones on
the left-hand side are not.
Recalling that a denotes the dilation operator (see Section 2 for definition), we
first estimate from above the right-hand side of (5.6):2nk=1
fk(s)xk(t)
X
=
2nk=1
fk(tk)xk(t0)
X
=
2nk=1
[0,2n](tk)xk(t0)
X
=
2nk=1
2n(xk)X
2nk=1
2n(x2n)
X
(since xk xk+1)
= x2nX (since the 2n(x2n)s are disjointly supported)
=
2nn2ni=1
gi
X
(from the definition of x2n)
2 (from (5.7)).
Consider the left-hand side of (5.6). We show that2nk=1 fk(s)xk(t)
X(II)is
unbounded as n . Since xk xk+1 for k = 1, . . . , 2n 1, then
(5.8)
2nk=1
fk(s)xk(t)
X(II)
2nk=1
fk(s)x1(t)
X(II)
.
Hence, we first estimate2n
k=1 fk(s) from below.
-
8/3/2019 S.V. Astashkin, F.A. Sukochev and C.P. Wong- Disjointification of Martingale Differences and Conditionally Independ
21/32
21
Since2n!
(2n n)! (2n
n)n and 1
1
2n 1
n
2n,
then, if n is sufficiently large, we have that
s [0, 1] :
2nk=1
fk(s) n 2
n!
n!(2n n)!
1
2n
n1 1
2n
2nn
(2n n)n
n!(2n)n (2
n n)2nn(2n)2
nn=
1
n!
1 n
2n
2n 1
n2n.
Therefore, if A :=
s [0, 1] : 2nk=1 fk(s) n , then by (5.8)
2n
k=1
fk(s)xk(t)X(II)
nA(s)x1(t)
X(II) = n
n2n(x1)
X .
But by definition of x1 and (5.7), n2n(x1)X = g1X 1/2. Hence, takinginto account the previous inequality, we obtain that
2nk=1
fk(s)xk(t)
X(II)
as n .
Thus, the left-hand side of (5.6) is unbounded as n and the result follows.
Remark 5.6. Assumption (5.1) for a sequence of independent functions {fk}nk=1 X is essential. In fact, Theorems 3.1 and 3.4 show that the inequality n
i=1
ri(t)xi(s)X(II)
C n
i=1
xi
Z2X
,
where {ri(t)} are the Rademacher functions, holds for any sequence {xi}ni=1 Xonly if X L2.
Remark 5.7. Recall that some disjointification relations using the space Z1X instead
of Z2X were obtained earlier for nonnegative independent random variables as well
(see [21, 6, 8]). However, in the case of mds and even conditionally independent
functions analogous formulas do not hold for any other symmetric spaces apart from
L1.
In fact, suppose that a symmetric space X satisfies the following condition: there
is a constant C > 0 such that for any sequence of independent functions {fk}nk=1 X and any sequence of measurable functions {xk}nk=1 X
(5.9) n
i=1
fi(s)xi(t)X(II)
C n
i=1
fi(s)xi(t)Z1X
.
-
8/3/2019 S.V. Astashkin, F.A. Sukochev and C.P. Wong- Disjointification of Martingale Differences and Conditionally Independ
22/32
22 S. V. ASTASHKIN, F. A. SUKOCHEV, AND C. P. WONG
Take fk(s) = 1 and xk(t) = [0,1/n](t) (k = 1, 2, . . . , n). Thenn
k=1 fk(s)xk(t) =
n[0,1/n](t) and the sumnk=1
fk(s)xk(t) is equimeasurable with [0,1]. Then in-
equality (5.9) gives
n[0,1/n]X C[0,1]X .
At the same time,
[0,1]X =
nk=1
[0,1/n]
X
n
k=1
[0,1/n]X = n[0,1/n]X .Hence, we have that
(5.10)[0,1/n]X 1n [0,1]X (n N).
Next, recall that given the symmetric space X, the function X(t) := AX ,where (A) = t, is the fundamental function of X. Recall also that X X,where X is the corresponding Lorentz space (see Section 2). It is not hard to
check that (5.10) implies: X(t) t. Therefore, X = L1, whence L1 X. Butfor all symmetric spaces X on [0, 1] we have that X L1. Thus, inequality (5.9)holds only in the case X = L1.
6. Comparison of martingale differences and nonnegative random
variables with their independent copies
The results on the disjointification of martingale differences proven in the previ-
ous sections allow us to compare norms of sums of martingale differences with the
norms of sums of their independent copies.
Theorem 6.1. Let X be a symmetric space on [0, 1], X I(L1, L2). Then thereexists a constant C > 0 such that for every mds {dk}nk=1 X and any sequence
{fk}n
k=1 X of independent functions such that fk is equimeasurable with dk forevery k = 1, 2, . . . , n we have that
(6.1) max
k=1,2,...,n
ki=1
di
X C
nk=1
fk
X
.
Proof. Since1
0fk(t) dt =
10
dk(t) dt = 0 (k = 1, 2, . . . , n), then {fk}nk=1 is a se-quence of mean zero independent functions. Therefore, by [21, Theorem 1], there
-
8/3/2019 S.V. Astashkin, F.A. Sukochev and C.P. Wong- Disjointification of Martingale Differences and Conditionally Independ
23/32
23
is a constant C > 0, depending only on X, such that
nk=1
fkZ2X
Cn
k=1fkX .
Combining this inequality with inequality (3.2) proved in Corollary 3.2 and applying
the fact that the functionsn
k=1 dk andn
k=1 fk are equimeasurable, we get the
result.
Theorem 6.2. Suppose that a symmetric space X I(L2, L) has order semi-continuous norm and qX < . Then there exists a constant C > 0 such that forevery mds {dk}nk=1 X and any sequence {fk}nk=1 X of independent functions
such that fk is equimeasurable with dk for every k = 1, 2, . . . , n we have that
(6.2) n
k=1
fk
X C
maxk=1,2,...,n
ki=1
di
X
.
Proof. Since qX < , the operator K is bounded in X. So we may apply [7,Theorem 3.1] to find a constant C > 0, depending only on X, such that n
k=1
fk
X C
nk=1
fk
Z2X
.
Inequality (6.2) follows from this and inequality (3.7) from Corollary 3.6.
For sequences of conditionally independent random variables {fk(s)xk(t)}nk=1satisfying condition (5.1) a rather stronger result holds. From Theorem 5.3 and
[21, Theorem 1] it follows
Theorem 6.3. Let X be a symmetric space on [0, 1] that either has the Fatou
property or is separable. Suppose qX < . Then there exists a constant C > 0such that for arbitrary sequences of independent functions {fk}nk=1 X satisfyingcondition (5.1), measurable functions {xk}nk=1 X and for any sequence of inde-pendent functions {gk}nk=1 X such that gk is equimeasurable with the functionfk(s)xk(t) (k = 1, 2, . . . , n), we have that
(6.3)
n
k=1
fk(s)xk(t)
X(II)
C
nk=1
gk
X
.
Remark 6.4. LetX be a symmetric space on [0, 1] that either has the Fatou prop-
erty or is separable. Suppose also that the operatorK : X X. Then, by [7,Theorem 3.1] and Remark 5.4, we obtain the converse of inequality (6.3), i.e., if
-
8/3/2019 S.V. Astashkin, F.A. Sukochev and C.P. Wong- Disjointification of Martingale Differences and Conditionally Independ
24/32
24 S. V. ASTASHKIN, F. A. SUKOCHEV, AND C. P. WONG
sequences {fk}nk=1 X, {xk}nk=1 X, and {gk}nk=1 X satisfy the same condi-tions as in Theorem 6.3, then
(6.4)
n
k=1
gk
X
C
nk=1
fk(s)xk(t)
X(II)
for some constantC > 0.
Now, we prove some inequalities comparing sums of nonnegative random vari-
ables to sums of their independent copies in the setting of symmetric spaces. These
inequalities are strongly related to the results of Theorems 6.1 and 6.2. Further-
more, the results on nonnegative random variables and their independent copies
provide a simpler method of obtaining the modular comparison inequalities (see
[28, p. 82] for terminology) proved in [17] (see also [18]).
Theorem 6.5. Suppose the operatorK is bounded on a symmetric space X on[0, 1]that either has the Fatou property or is separable. Then for arbitrary sequences of
nonnegative functions {gk}nk=1 and independent functions {fk}nk=1 from X suchthat fk is equimeasurable with gk for every k = 1, 2, . . . , n we have that
(6.5)
nk=1
fk
X
KXX
nk=1
gk
X
,
with a universal constant > 0.
Proof. First, let Z[0,) be a symmetric space on the semiaxis [0,) and{hk}nk=1 Z[0,) (n N) be an arbitrary sequence of nonnegative measurablefunctions on [0,) whose supports have finite measure. Obviously, for every > 0,
0
nk=1
hk
(s) ds
0
nk=1
hk
(s) ds.
Then if Z[0,) I(L1[0,), L[0,)), by [23, Theorem 2.4.3], we obtain that
(6.6) n
k=1hkZ[0,) C
nk=1
hkZ[0,),
where C is the interpolation constant of the space Z[0,) in the couple(L1[0,), L[0,)).
Note that the space Z1X also either has the Fatou property or is separable (de-
pending on X). Therefore, by assumption, Z1X I(L1[0,), L[0,)) with con-stant 1.
-
8/3/2019 S.V. Astashkin, F.A. Sukochev and C.P. Wong- Disjointification of Martingale Differences and Conditionally Independ
25/32
25
Now, let sequences {gk}nk=1 and {fk}nk=1 X satisfy our assumptions. Then,applying arguments similar to that at the beginning of the proof together with
inequality (6.6) to the space Z1X , we obtain that
(6.7) n
k=1
fk
Z1X
n
k=1
gk
X
.
On the other hand, by [8, Theorem 1(ii)], there is a universal constant > 0 such
that nk=1
fk
X KXX
nk=1
fk
Z1X
.
Combining this inequality with (6.7), we obtain (6.5).
Remark 6.6. Using a different (noninterpolation) method, inequality (6.6) in the
case when a symmetric space Z[0,) is separable, is proved in [19, Lemma 7.2].
Now we are able to prove the following modular inequality.
Corollary 6.7. Let M be an Orlicz function on [0,) such that the operatorK is bounded on the Orlicz space LM = LM[0, 1]. Then there exists a constantC > 0 such that for every n N and for arbitrary sequences of nonnegative mea-
surable functions {gk}n
k=1 and independent functions {fk}n
k=1 from X such that fkis equimeasurable with gk (k = 1, 2, . . . , n), there is a > 0, depending only on
nk=1 gkLM , such that(6.8)
10
M
nk=1 fk(s)
ds
10
M
Cn
k=1 gk(s)
ds.
In particular, if nk=1 gkLM = 1, we may take = C to obtain:(6.9)
10
M
nk=1 fk(s)
C
ds
10
M
n
k=1
gk(s)
ds.
Proof. Since an arbitrary Orlicz space has the Fatou property, from Theorem 6.5,
it follows that nk=1
fk
LM
C nk=1
gk
LM
,
where C := KLMLM . If := Cn
k=1 gkLM , then10
M
Cn
k=1 gk(s)
ds = 1.
-
8/3/2019 S.V. Astashkin, F.A. Sukochev and C.P. Wong- Disjointification of Martingale Differences and Conditionally Independ
26/32
26 S. V. ASTASHKIN, F. A. SUKOCHEV, AND C. P. WONG
Hence, from the previous inequality and the definition of the Luxemburg norm in
LM we have that 10
M
nk=1 fk(s)
ds 1,
which implies (6.8). Inequality (6.9) follows from (6.8) and the definition of .
Let us show how inequalities for nonnegative random variables and their in-
dependent copies imply analogous inequalities for martingale differences and their
independent copies. But first, we need an auxiliary result (see e.g. [5, Corollary 15]).
Recall that a Banach lattice E is said to be pconvex (1 p < ) with constantK
1 if n
k=1
|xk|p1/p K n
k=1
||xk||p1/p
for every choice of vectors x1, x2, . . . , xn from E. For a pconvex Banach function
lattice E we may define its pconcavification, i.e., the Banach lattice E(p) with the
norm xE(p) := |x|1/ppE .
Lemma 6.8. Suppose that p > 1 and the Kruglov operatorK is bounded on a pconvex symmetric space X. ThenK is bounded onX(p) as well andKX(p)X(p)
Kp
XX.
Proof. By definition, Kf is equimeasurable with the functionn=1
nk=1
fn,k(x)En (x), x [0, 1],
where {En} is a sequence of measurable pairwise disjoint subsets of [0, 1], (En) =1/(en!) (n N), and fn,1, fn,2, . . . , f n,n are copies of f such that the sequencefn,1, fn,2, . . . , f n,n, En consists of independent functions. Then, by the elementary
inequality nk=1
k
1/p nk=1
|k|1/p (k R),
we see that (Kf)1/p (K(|f|1/p)). Therefore,
KfX(p) = |Kf|1/ppX K(|f|1/p)pX KpXXfX(p) ,
and we obtain the result.
-
8/3/2019 S.V. Astashkin, F.A. Sukochev and C.P. Wong- Disjointification of Martingale Differences and Conditionally Independ
27/32
27
Theorem 6.9. Suppose that the operatorK is bounded in a 2convex symmetricspace X on [0, 1] that either has the Fatou property or is separable. Then for an
arbitrary mds {dk}nk=1 X and for any sequence of independent functions {fk}nk=1such that fk is equimeasurable with dk for every k = 1, 2, . . . , n we have that
(6.10) n
k=1
fk
X C1
nk=1
d2k
1/2X
,
with a constant C1 > 0.
In particular, if qX < , then there is a constant C2 > 0 such that
(6.11) n
k=1
fkX C2 maxk=1,2,...,n k
i=1
diX .Proof. Note that the space X(2) has the Fatou property (respectively, is separable) if
the space X has the Fatou property (respectively, is separable). Therefore, applying
Lemma 6.8 and Theorem 6.5 for sequences {d2k}nk=1 and {f2k}nk=1, we obtain that nk=1
f2k
X(2)
K2XX n
k=1
d2k
X(2)
,
which implies that
n
k=1
f2k1/2X 1/2KXX
n
k=1
d2k1/2
X.
Since {fk}nk=1 is a sequence of mean zero independent functions and K is boundedin X, then, by [2, Theorem 1],
nk=1
f2k
1/2X nk=1
fk
X
.
This and the previous inequality imply (6.10). Assuming qX < , we get inequality(6.11) as an immediate consequence of (6.10) and [20, Theorem 3].
Remark 6.10. In Theorem 6.2, inequality (6.11) was proved under a weaker as-
sumption X I(L2, L).
Theorems 6.1, 6.5 and 6.9 can be viewed as extensions of modular inequalities
proved in [17] and [18] to the setting of symmetric spaces. On the other hand, using
results obtained here, we can prove these modular inequalities. Let us prove, for
instance, the modular version of (6.5).
-
8/3/2019 S.V. Astashkin, F.A. Sukochev and C.P. Wong- Disjointification of Martingale Differences and Conditionally Independ
28/32
28 S. V. ASTASHKIN, F. A. SUKOCHEV, AND C. P. WONG
In the following, we will use the notation EF =1
0F(s) ds. Recall that an Orlicz
function M on [0,
) satisfies the 2condition at infinity if there exist C > 0 and
u0 > 0 such that M(2u) CM(u) for all u u0.
Theorem 6.11. LetM be an Orlicz function on [0,) satisfying the 2conditionat infinity. Then, possibly after changing the function M on the segment [0, 1], there
exists a constant C > 0 such that for every n N and for arbitrary sequences ofnonnegative measurable functions {gk}nk=1 and independent functions {fk}nk=1 fromthe Orlicz space LM on[0, 1] such thatfk is equimeasurable withgk (k = 1, 2, . . . , n),
we have that
(6.12) E
M n
k=1
fk CEM n
k=1
gk
.
Proof. First of all, note that by assumption, the upper Boyd index qLM of LM is
finite (see e.g. [26, Theorem 11.7]). Therefore, since qLMt = qLM < , whereMt :=
1t M (t > 0), then K : LMt LMt . Moreover, for any t > 0, LMt has the
Fatou property. Therefore, by Theorem 6.5,
(6.13)
nk=1
fk
LMt
KLMtLMt
nk=1
gk
LMt
,
where > 0 is a universal constant. Let us show that
(6.14) supt>0
KLMtLMt C1KLMLM
with some C1 > 0.
Without loss of generality, we may (and shall) assume that
(6.15) M(u + v) (M(u) + M(v))
for some constant > 0 and all u, v 0 (see for e.g. [14, Formula (7.9)] or [25,Proposition 2.b.5]).
Let now f 0. Suppose that EM(f) := A < . If {fk}k=1 is a sequence ofindependent functions equimeasurable with f, then, by (6.15), for every n N
E
M n
k=1
fk
E
M n1
k=1
fk
+ M(fn)
= EM
n1k=1
fk
+ A
2EM
n2k=1
fk
+ 2A n1nA.
-
8/3/2019 S.V. Astashkin, F.A. Sukochev and C.P. Wong- Disjointification of Martingale Differences and Conditionally Independ
29/32
29
By definition, Kf equals nk=1
fk on a set En with Lebesgue measure (En) =
1/(en!) (n = 1, 2, . . . ); moreover, En are pairwise disjoint and the family
{f1, . . . , f n, En} consists of independent functions. Therefore,
EM(Kf) =n=1
EM n
k=1
fkEn
=
n=1
EM n
k=1
fk
(En)
n=1
n1nA(En) =A
e
n=1
n1
(n 1)! = e1A,
whence,
EMt(Kf) e1EMt(f) (t > 0).From this it follows that
KfLMt e1fLMt (t > 0),
i.e., (6.14) is proved. Therefore, by (6.13), there is a constant C > 0 such that
(6.16) n
k=1
fk
LMt
C nk=1
gk
LMt
(t > 0),
for every n N, where {gk}nk=1 is an arbitrary sequence of nonnegative measurablefunctions from LM and {fk}nk=1 is any sequence of their independent copies.
Furthermore, we may find t > 0 such that
(6.17)
1
0
M
Cnk=1
gk(s)
ds = t,
or equivalently, nk=1
gk
LMt
=1
C.
Then from (6.16) it follows that10
M
Cn
k=1
fk(s)
ds t.
Combining this with equality (6.17), we obtain that10
M n
k=1
fk(s)
ds 1
0
M
Cn
k=1
gk(s)
ds
and the proof is complete because M satisfies the 2condition at infinity.
Using Theorem 6.9, arguments analogous to the ones used in the proof of the
previous theorem and the BurkholderDavisGandy square function inequality from
[15], it is not hard to prove the following comparison assertion.
-
8/3/2019 S.V. Astashkin, F.A. Sukochev and C.P. Wong- Disjointification of Martingale Differences and Conditionally Independ
30/32
30 S. V. ASTASHKIN, F. A. SUKOCHEV, AND C. P. WONG
Theorem 6.12. LetM be an Orlicz function on [0,) satisfying the 2conditionat infinity such that the function M(
t) is convex for t > 0. Then, possibly after
changing the function M on the segment [0, 1], there exists a constant C > 0 such
that for an arbitrary mds {dk}nk=1 on [0, 1] and for any sequence of independentfunctions {fk}nk=1 such that fk is equimeasurable with dk (k = 1, 2, . . . , n), we have
(6.18) E
M n
k=1
fk
CE
M
maxk=1,2,...,n
ki=1
di
.Remark 6.13. Similar disjointification methods allow us to prove the converse of
the modular inequalities (6.12) and (6.18) (see [17] and [18]) as well. We wont
pursue this subject here, noting only that relation (6.1) proved in Theorem 6.1 is
an extension of these inequalities to the setting of symmetric spaces.
References
[1] C. D. Aliprantis and O. Burkinshaw, Positive operators, Academic Press, Florida, 1985.
[2] S. V. Astashkin, Independent functions in symmetric spaces and Kruglov property, Mat.
Sbornik 199 (2008), no. 7, 320 (in Russian); English transl. in Sbornik: Math. 199 (2008),
no. 7, 945963.
[3] S. V. Astashkin and L. Maligranda, Interpolation between L1 and Lp, 1 < p < , Proc.
Amer. Math. Soc., 132 (2004), no. 10, 29292938.
[4] S. V. Astashkin, E. M. Semenov, and F. A. Sukochev,The Banach-Saks p-property
, Math.Ann., 332 (2005), no. 4, 879900.
[5] S. V. Astashkin, E. M. Semenov, F. A. Sukochev and D.V. Zanin, Kruglov operator and
operators defined by random permutations, Funktsional. Anal. i Prilozh. 43(2009), no. 2,
318 (in Russian); English transl. in Funct. Anal. Appl. 43 (2009), no. 2, 8395.
[6] S. V. Astashkin and F. A. Sukochev, Sums of independent random variables in symmetric
spaces: an operator approach. Israel J. Math., 145 (2005), 125156.
[7] S. V. Astashkin and F. A. Sukochev, Series of independent mean zero random variables in
symmetric spaces with the Kruglov property, Zap. nauchn. semin. POMI. Issled. po lin. oper.
i teor. f-ii. Ch. 35. 345 (2007), 2550 (in Russian); English transl. in J. Math. Sci. 148 (2008),
no. 6, 795809.[8] S. V. Astashkin and F. A. Sukochev, Best constants in Rosenthaltype inequalities and the
Kruglov operator, Ann. Probab., 38 (2010), no. 5, 19862008.
[9] S. V. Astashkin and F. A. Sukochev, Independent functions and Banach space geometry,
Uspekhi matemat. nauk 65 (2010), no. 6, 386 (in Russian); English transl. in Russian
Math. Surveys, 65 (2010), no. 6, 10031081.
[10] S. V. Astashkin, On finite representability of lpspaces in symmetric spaces, Algebra i Analiz,
23 (2011), no. 2, 77101 (in Russian).
-
8/3/2019 S.V. Astashkin, F.A. Sukochev and C.P. Wong- Disjointification of Martingale Differences and Conditionally Independ
31/32
31
[11] C. Bennett and R. Sharpley, Interpolation of Operators, Academic Press, Inc., Orlando,
Florida, 1988.
[12] M. Sh. Braverman, Independent random variables and symmetric spaces, Cambridge Uni-
versity Press, Melbourne, 1994.
[13] Ju. A. Brudnyi and N. Ya. Krugljak, Interpolation Functors and Interpolation Spaces,Vol. I,
North-Holland Publishing Co., Amsterdam, 1991.
[14] D.L. Burkholder, Distribution Function Inequalities for Martingales Ann. Probab., 1 (1973),
no. 1, 1942.
[15] D. L. Burkholder, B. J. Davis, and R. F. Gandy, Integral inequalities for convex function of
operators on martingales, in Proc. Sixth Berkeley Symp. Math. Statist. Prob. 2, (1972),
223240.
[16] N.L. Carothers and S. J. Dilworth, Equidistributed random variables in Lp,q, J. Funct. Anal.,
84 (1989), 146159.
[17] V. H. de la Pena, Bounds on the expectation of functions of martingales and sums of positive
rvs in terms of norms of sums of independent random variables, Proc. Amer. Math. Soc.,
108 (1989), no. 1, 233239.
[18] P. Hitczenko, A remark on the paper Martingale inequalities in symmetric function spaces
by W.B. Johnson and G. Schechtman, Functional Analysis-Lecture Notes in Mathematics,
Springer, Berlin, 1991, 1470, 177182.
[19] W. B. Johnson, B. Maurey, G. Schechtman and L. Tzafriri, Symmetric structures in Banach
spaces, Mem. Amer. Math. Soc., 1979. V. 1, no. 217.
[20] W. B. Johnson, G. Schechtman, Martingale inequalities in rearrangement invariant function
spaces, Israel J. Math., 64 (1988), 267275.
[21] W. B. Johnson and G. Schechtman, Sums of independent random variables in symmetric
function spaces, Ann. Probab., 17 (1989), 789808.
[22] L. V. Kantorovich and G. P. Akilov, Functional Analysis, Nauka, Moscow 1977; English
transl. Pergamon Press, Oxford-Elmsford, New York 1982.
[23] S.G. Krein, Ju.I. Petunin and E.M. Semenov, Interpolation of linear operators, Translations
of Mathematical Monographs, Amer. Math. Soc. 54 (1982).
[24] V.M. Kruglov, A remark on infinitely divisible distributions, Teor. Ver. Pril., 15 (1970),
331336 (Russian).
[25] J. Lindenstrauss and L. Tzafriri, Classical Banach Spaces vol. II, Springer-Verlag, Berlin,
1979.
[26] L. Maligranda, Orlicz Spaces and Interpolation, Seminars in Math. 5, Univ. of Campinas,
Campinas SP, Brazil 1989.
[27] M. Mastylo, Interpolation of linear operators in the Kothe dual spaces, Ann. Mat. Pura Appl.
(IV) 154 (1989), 231242.
[28] M.M. Rao and Z.D. Ren, Theory of Orlicz spaces Marcel Dekker, Inc., New York, 1991.
-
8/3/2019 S.V. Astashkin, F.A. Sukochev and C.P. Wong- Disjointification of Martingale Differences and Conditionally Independ
32/32
32 S. V. ASTASHKIN, F. A. SUKOCHEV, AND C. P. WONG
[29] H. P. Rosenthal, On the subspaces of Lp (p > 2) spanned by sequences of independent random
variables, Isr. J. Math., 8 (1970), 273303.
[30] A. N. Shiryaev, Probability, Springer-Verlag, New York, 1984.
S.V.A.: Department of Mathematics, Samara State University, Samara, Russia
E-mail address: [email protected]
F.A.S.: School of Mathematics and Statistics, University of New South Wales, Syd-
ney NSW 2052, Australia
E-mail address: [email protected]
C.P.W.: School of Mathematics and Statistics, University of New South Wales,
Sydney NSW 2052, Australia